nur-dev/frozen-stress-r0-blsp

BTA — R0 BLSP cohort (5 seeds)

5-seed cohort of the BLSP-class baseline adapter for the Beyond Transcript Alignment research project.

Architecture: Conv1d(1024→4096, k=4, s=4) → MLP-2 → RMSNorm → +modality token (84M trainable params) Encoder: WavLM-Large layer 16 (microsoft/wavlm-large, frozen) LLM: Qwen3-8B-Instruct (Qwen/Qwen3-8B, frozen, enable_thinking=False) Loss: $\mathcal{L}{\mathrm{BLSP}} = \mathcal{L}{\mathrm{task}} + \lambda_{\mathrm{KL}} \mathcal{L}_{\mathrm{KL}}$ Optimizer: AdamW lr=5e-5, weight_decay=0.01, max_norm=1.0; cosine to 1e-6 Schedule: seed 1234 = 600 steps + warmup 500; seeds 1235-1238 = 300 steps + warmup 250 Code: https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress

Files

File	Size
A_BLSP_seed1234.pt	336 MB
A_BLSP_seed1235.pt	336 MB
A_BLSP_seed1236.pt	336 MB
A_BLSP_seed1237.pt	336 MB
A_BLSP_seed1238.pt	336 MB
adapter_init.json	calibrated init (std_8B=0.02205, RMSNorm scale=0.0215)

Reported metrics (5-seed cohort)

Metric	Mean ± σ
Probe-G total	0.6306 ± 0.0095
Probe-G$_{\mathrm{neutral}}$	0.5122 ± 0.0108
Probe-G$_{\mathrm{explicit}}$	0.7491 ± 0.0120
Probe-K linear eval_full	0.2105 ± 0.0128
Probe-K MLP-2 eval_full	0.2446 ± 0.0339

Sub-cascade collapse: every seed's Probe-K linear is below the $K_T$ text-only baseline of 0.290.

Usage

import torch
from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("nur-dev/frozen-stress-r0-blsp", "A_BLSP_seed1234.pt")
state = torch.load(ckpt, map_location="cpu")
# Adapter loading — see scripts/stage2_eval.py in the GitHub repo

License

CC-BY-NC-4.0 (weights inherit non-commercial restriction from the training datasets Stress-17K-raw, StressPresso, Expresso). Permitted for academic research and ablation studies.